Artificial Intelligence as a Disruptor to AMD Diagnosis

Companies plan for an expanded role for AI in retina therapies.

Not long from now, patients may come to you with a referral from an artificial intelligence (AI) system. The patient will bring — or perhaps send — images taken by a technician at a drug store with the guidance of another AI. You will use a third AI to plan these patients’ treatments for age-related macular degeneration (AMD).

As futuristic as this sounds, the technology already exists. “We are delighted to say that the tool is now ready for clinical use and potential remote telemedical deployment,” said Sharmina Alauddin, MBBS, describing a program developed by iHealthScreen at the 2002 annual meeting of the Association for Research in Vision and Ophthalmology.1

iHealthScreen hopes to emulate the success of IDx-DR, the first-ever autonomous AI for detecting diabetic retinopathy (DR) approved by the US Food and Drug Administration (FDA). Several primary care and university clinics are using IDx-DR to screen patients for DR and refer them to retina specialists. The company sees a huge potential for expansion, and now it has its eye on the market for an AMD device. Many other systems are advancing toward the marketplace. DeepMind, a project of Google, has dived into AMD assessment in collaboration with Moorfields Eye Hospital in London. And researchers at universities in the United States, Europe, and Asia have churned out papers in recent years exploring various similar approaches.

“It’s an important subject and it’s exploding now because of the of the COVID-19 epidemic,” said R. Theodore Smith, MD, PhD, a professor of ophthalmology and neuroscience at New York Eye and Ear Infirmary of Mount Sinai Hospital in New York. “It’s making us rethink how we do our imaging. The patients could almost take their own images and send them to you.”

AN UNMET NEED

Even before the pandemic, Smith argued, AI could meet a pressing need. A leading cause of blindness, AMD afflicts more than 8 million people in the United States.2

Catching the condition early can make a key difference. The AREDS formulation of nutritional supplements can slow its progress. Anti-vascular endothelial growth factor (VEGF) drugs can treat disease if it reaches the neovascular stage, and some research is under way to see if they can have a prophylactic effect. A variety of treatments under study have shown potential for preserving vision in dry AMD, including photobiomodulation and subthreshold nanosecond laser.3

But these treatments can only prevent future decline. By the time they are diagnosed, many patients will have lost vision that can’t be restored. There aren’t enough retina specialists to screen that many people. Even if there were, some people live far from retina centers or can’t afford the cost of a screening.

That’s where AI comes in.

Applications for AI in AMD fall into 3 broad categories. First comes the AI used to guide the technician in handling a fundus camera. IDx has deployed such as system in its DR program. The operator needs only a high school education to use the “robotic” camera, says IDx founder Michael Abramoff, MD, PhD. Next comes the AI that analyzes the images to screen patients for the presence of the disease. And finally comes AI that provides a prognosis, helping physicians determine which patients to follow most closely and when to intervene.

SCREENING

Researchers are experimenting with a range of designs for AI to screen for AMD. All of them share a fundamental principle: unlike conventional computer programs, AI can learn by experience. Given a set of labeled images, the computer looks for patterns that distinguish them so it can recognize similar images when it sees them again.

Programmers have drawn on a variety of inspirations to create these programs, including the way ant colonies figure out the shortest path to a food source.4 Deep learning, the hottest trend in unsupervised learning, simulates the neural networks in a human brain. Whenever a factor proves predictive, the AI assigns more weight to the connection between neurons that represents that factor. Successive layers of networks identify finer and finer distinctions.

In one approach, programmers label specific features in each image in the training data set, such as pigmentation, exudate, and hemorrhage or the number, size, and shape of drusen. The computer learns to recognize them and to determine how these correlate with a stage of AMD (including no AMD at all). Given a fresh set of images, the computer finds features of AMD, then applies its algorithm to determine which images show retinas that should be evaluated by a human specialist. That’s the approach that IDx is taking in its AMD technology.

In another approach, the programmers simply label images by their stage of AMD (including no AMD). Then, based on the pixels in the images, the AI system finds its own patterns that distinguish the different disease conditions. In this approach, the AI can identify characteristics unknown to the programmer. In some of the most dramatic — though not particularly useful — demonstrations of this potential, an AI system was able to distinguish retinas from female subjects from retinas of male subjects.5 No one has been able to figure out what feature the computer used to make that distinction.

But this advantage can also be a disadvantage. Since the criteria the AI is using remains in an undecipherable “black box,” it may include irrelevant data. “It might take a part of the image that are outside the retina that’s coming from the camera,” says Abramoff. “If I have training data from 2 different clinics, let’s say it’s slightly asymmetric in one clinic, slightly shifted to the left. Now if it happens that this clinic has more patients with AMD than another, the machine will start to associate the asymmetry with the disease.”

Programmers have also created algorithms to predict AMD progression using patients’ genes or demographics such as age, race, sex, diabetes, body mass index, visual acuity, and sunlight exposure. Some systems combine multiple approaches.3

Many such AI systems have achieved high accuracy in categorizing fundus images by stage of AMD. But the programmers so far have mostly tested their AI systems on the same data sets used to train them. It’s not clear how well these AI systems would perform in a clinic where patients have not been selected for research, or where the cameras used are different from the ones used for the data set.

iHealthScreen programmers say they are the first to have met the challenge of real-world screening with a trial in 160 unselected nondilated subjects aged over 50 years. Comparing the system’s results with those of 2 ophthalmologists, it achieved 88.67% accuracy with sensitivity (true positive) of 86.57% and specificity (true negative) of 90.36%.1

PREDICTION

If such AI systems begin referring millions of new patients to retina practices, how will retina specialists keep up with all of them? AI can help there also, by analyzing fundus or optical coherence tomography (OCT) images to predict which eyes are most likely to progress to more serious stages of disease. Potentially, such systems could guide the timing of anti-VEGF injections and other treatments.

“If we could identify those patients, we could at a minimum monitor those patients more closely,” says Pearse Keane, MD, MSc, an associate professor of ophthalmology and brain sciences at University College, London. In collaboration, with DeepMind, Pearse and his colleagues programmed an AI system that uses OCT images to predict which fellow eyes of eyes with AMD were likely to develop the disease.6 Others have shown success in distinguishing OCT scans of eyes with exudative changes from those without such changes.7

How soon will this technology start affecting your practice? Programs used by retina specialists to help monitor their patients might arrive soon. Such “assistive” AI requires much less scrutiny from the FDA than the autonomous screening AI systems meant to operate in the absence of a licensed physicians. But iHealthScreen is already in talks with the agency about the design of a larger clinical trial of its screening program.

IDx, too, has plans for an AMD screening trial — once the company is satisfied that it can work out reimbursement and liability questions. “Too many people don’t get the health care that they need,” Abramoff says. “You can solve these problem with AI but if you do it wrong then you get a backlash. That is my concern.”

Once such hurdles are overcome, the potential for AI in retinal imaging is huge — and not just for DR and AMD. “I’m particularly interested in quantifying drusen, and not just for staging early and intermediate AMD, but also as a new biomarker for other diseases,” says Emma Pead, PhD, a research fellow at the University of Edinburgh in the United Kingdom.

Correlations have surfaced between retinal features and glaucoma, cardiovascular disease, and Alzheimer disease.4 One day, Pead and others hope to be able to diagnose these conditions through the retina, observe how treatments affect them, or discover new insights into the way the diseases progress. “It’s not an easy task, but we’re getting there,” she said. RP

Editor’s note: This article is discussed on the Retina Podcast. Listen at www.retinapodcast.com.

REFERENCES
  1. Alauddin S, Bhuiyan A, Govindaiah A. A prospective trial of an artificial intelligence based telemedicine platform to stratify severity of age-related macular degeneration (AMD). Invest Ophthalmol Vis Sci. 2020;61:1843.
  2. Bressler NM, Bressler SB, Congdon NG, et al. Potential public health impact of Age-Related Eye Disease Study results: AREDS report no. 11. Arch Ophthalmol. 2003;121(11):1621-1624.
  3. Bhuiyan A, Wong TY, Ting DSW, Govindaiah A, Souied EH, Smith RT. Artificial Intelligence to Stratify Severity of Age-Related Macular Degeneration (AMD) and Predict Risk of Progression to Late AMD. Transl Vis Sci Technol. 2020;9(2):25.
  4. Pead E, Megaw R, Cameron J, et al. Automated detection of age-related macular degeneration in color fundus photography: a systematic review. Surv Ophthalmol. 2019;64(4):498-511.
  5. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018;2(3):158-164.
  6. Yim J, Chopra R, Spitz T, et al. Predicting conversion to wet age-related macular degeneration using deep learning. Nat Med. 2020;26(6):892-899.
  7. Motozawa N, An G, Takagi S, et al. Optical coherence tomography-based deep-learning models for classifying normal and age-related macular degeneration and exudative and non-exudative age-related macular degeneration changes. Ophthalmol Ther. 2019;8(4):527-539.